Welcome to Phase Processing Lab conducted by Pejman Mowlaee CV

On the Importance of Harmonic Phase Modification For Improved Speech Signal Reconstruction

- Audio samples -




- Experiment 1 (Proof-of-concept) -


Female speaker in white noise at SNR = 10 (dB)


We consider different combinations for decomposed phase: linear and unwrapped phase, depending on their noisy, oracle (clean), or estimated versions. Here, we aim for justifying the hypothesis if an oracle linear and unwrapped phase suffice, rather than a clean STFT phase, for an improved signal reconstruction, when combined with an amplitude-only enhancement.



Figure below, demonstrates the proof-of-concept results; (top) spectrogram, (middle) group delay, (bottom) phase variance. The results, from left to right, are shown for the following phase scenarios: clean signal (clean phase), oracle linear phase + oracle unwrapped phase, oracle linear phase + noisy unwrapped phase, noisy linear phase + oracle unwrapped phase, noisy linear phase + noisy unwrapped phase, noisy phase. The noisy signal is produced by contaminating a female utterance saying "been blue at L four soon" with white noise at SNR = 0 (dB). The audio wave files are presented below.
Noisy (Input) Clean (Reference)
Lin. Phase (Oracle) + Unwrapped (Oracle) Lin. Phase (Oracle) + Unwrapped (Noisy)
Lin. Phase (Est.) + Unwrapped (Oracle) Lin. Phase (Est.) + Unwrapped (Noisy)
Noisy STFT phase



- Experiment 2 -


Male Speaker in Babble noise at SNR = 5 (dB):


We aim for justifying the hypothesis if a proper linear phase + a proper unwrapped phase suffice for an improved signal reconstruction, combined with and amplitude-only enhancement. To this end, we compared the clean STFT phase versus the oracle linear + oracle unwrapped phase.
oracle lin, oracle unwrapped Clean STFT Phase